Berkeley Function-Calling Leaderboard — Vendored Sources
========================================================

The contents of this directory contain code and data adapted (and in some
cases copied verbatim) from the upstream Berkeley Function-Calling
Leaderboard (BFCL) project:

    https://github.com/ShishirPatil/gorilla
    Path: berkeley-function-call-leaderboard/bfcl_eval/

Upstream is licensed under the Apache License, Version 2.0.
See https://www.apache.org/licenses/LICENSE-2.0 for the full text.

Vendored modules and their upstream sources:

  * func_source_code/*.py
      <- bfcl_eval/eval_checker/multi_turn_eval/func_source_code/*.py
    Tool implementations (GorillaFileSystem, MathAPI, MessageAPI,
    TwitterAPI, TicketAPI, TradingBot, TravelAPI, VehicleControlAPI,
    WebSearchAPI, MemoryAPI_*).

  * memory_prereq_conversation/*.json
      <- bfcl_eval/data/memory_prereq_conversation/*.json
    Pre-baked conversation states used by the memory category prerequisite
    write-phase. Copied byte-for-byte.

  * memory_utils.py
      <- bfcl_eval/utils.py (memory-related helpers only) and
         bfcl_eval/eval_checker/agentic_eval/agentic_checker.py
    The subset of utils we need to drive the memory snapshot dance and
    score agentic responses (substring match w/ punctuation normalization).

  * _upstream_multi_turn_utils.py / _upstream_multi_turn_checker.py
      <- bfcl_eval/eval_checker/multi_turn_eval/multi_turn_*.py
    Kept as reference for the runtime/scoring logic we reimplemented in
    runtime.py (per-test instance scope instead of module globals).

Modifications from upstream:
  * Package-relative imports
    (``benchmarks.bfcl.executable_runtime.func_source_code...``).
  * Network-gated classes raise ``RuntimeNetworkRequired`` instead of
    failing opaquely when run without credentials.
  * Per-test ``ExecutableRuntime`` ownership of tool instances.

Apache-2.0 attribution is preserved on each vendored file.